16. Much of the output is biased, harmful, or unsafe

You’ll sound like a racist granddad who has learned how to write generic LinkedIn posts.

Summary

These AI systems are trained on the web. And, yes, that includes the porn, abuse, and violent imagery.

Many of these tools randomly output extremely distasteful content, some of which might even expose you to legal risk. The blocks vendors put into place to prevent this are often inadequate and often easily be bypassed. Removing unsafe data from the training set will in some cases also reduce the quality of its safe output, putting vendors in a position where they might have a higher tolerance for unsafe output than you.